An Intrinsic Hamiltonian Formulation of

Network Dynamics : Non-standard

Poisson Structures and Gyrators
by B. M. MASCHKE
Control Laboratory, Conservatoire National des Arts et M&tiers, 21 rue Pinel,
75013 Paris. France

A. J. VAN DER SCHAFT

Department of Applied Mathematics, University of Twente, P.O. Box 217,

7500 AE Enschede, The Netherlands

and P. c. BREEDVELD
Department of Electrical Engineering, University of Twente, P.O. Box 217,
7500 AE Enschede, The Netherlands

ABSTRACT : The aim of this paper is to provide an intrinsic Hamiltonian jormulation of the
equations of motion ofnetwork models of non-resistive physical systems. A recently developed
extension of the classical Hamiltonian equations of motion considers systems with state space
given by Poisson manifolds endowed with degenerate Poisson structures, examples of which
naturally appear in the reduction of’ systems with symmetry. The link with network
representations of non-resistive physical systems is established using the generalized bond
graph formalism which has the essential feature of symmetrizing all the energetic network
elements into a single class and introducing a coupling unit gyrator. The relation between the
Hamiltonian formalism and network dynamics is then investtgated throqh the representation
of the invariants of the system, either captured in the degeneracy of the Poisson structure or
in the topological constraints at the ports of the gyrative type network structure. This provides
a Hamiltonian formulation of dimension equal to the order of the physical system, in particular,
for odd dimensional systems. A striking example is the direct Hamiltonian formulation of
electrical LC networks.

I. Introduction

The modern mathematical theory of analytical mechanics evolved from the

Newtonian and variational formulations to a differential-geometric frame where
the Hamiltonian formalism is recognized as the fundamental axiom (14). Under-
lying the dynamic systems of analytical mechanics is a certain geometric structure,
called symplectic or Poisson structure of full-rank (5). Using this structure one

The Franklin Institute 001&0032/92 $5.00+0.00 923

B. M. Maschke et al.

gains crucial qualitative information, in particular on stability (2,4). Apart from

applications in mechanics, as in celestial mechanics (2) or mechanisms (6), this
geometric approach found applications in electrical engineering, for instance in
electrical circuit theory (7) as well as in control theory (8-10).
Arising from a different tradition based on network representations, a unified
formalization of physical systems emerged (11-13) covering domains as different
as chemical reactions (14,15), thermodynamics (13) and mechanisms (1618). It
displays an additional network structure on the fundamental physical concepts of
energy and conservation principles (13,19-22) where the constitutive assumptions
on the elements are based on a classification of physical variables and energies
derived in accordance with the fundaments of thermodynamics (2325).
This paper is intended to enlighten some aspects of physical systems theory in
the double context of analytical mechanics and network theory. The relation
between these two formalisms will be investigated through the different ways in
which the dynamical invariants of physical systems are expressed. The authors
regard this paper as the continuation of previous works on the structure of the
dynamics of network representations of electrical circuits (7,26), of chemical re-
actions (14,15), general non-dissipative systems (27) and control systems (8,9).
In Section II, the differential-geometric definition of Hamiltonian systems is
recalled in terms of a Poisson structure defined on the state-space manifold (5).
The main departure from a classical presentation is to allow for general Poisson
structures, including degenerate ones (28-32). The degeneracy of the Poisson struc-
ture is discussed with regard to invariance : a set of smooth real-valued functions
(different from the energy function) is left invariant by any Hamiltonian dynamics
defined with regard to the Poisson structure. Thus these invariants are fully captured
in the geometric structure of the state manifold. It is recalled that in contrast, for
standard Hamiltonian systems defined w.r.t. non-degenerate Poisson structures
(i.e. symplectic structures), the invariance properties are expressed as symmetries
of the Hamiltonian function. Finally, the Hamiltonian system defined on a state
manifold with degenerate Poisson structure will be regarded as the result of an
intermediate step of a reduction procedure from an embedding standard
Hamiltonian system with symmetries to a reduced standard Hamiltonian system.
Section III recalls the physical concepts underlying the modelling of physical
systems in network terms using the bond graph notation (11,12). More precisely
the generalized bond graph formalism (13) is used : its main feature is to symmetrize
all energetic elements to one class and to introduce a unit gyrator called “sym-
plectic” gyrator (25). which enables the relation with the Hamiltonian formalism.
Indeed the main objects of the Hamiltonian formalism (Hamiltonian function.
Poisson structure) will be related to a class of constitutive relations of some network
elements. Also the dynamic equations associated with a network model will be
shown to be Hamiltonian by allowing for degenerate Poisson structures on the
state-space manifold (this solves the problem of odd dimensional systems discussed
in (27,33,34)). The last part is devoted to the bond graph realization of the reduced
and embedding standard Hamiltonian systems in order to compare the different
representations of physical invariants in network as well as in differential geometric
terms.
 Humiltoniun Formulation of Network Dynumics

II. Poisson Structures, Symmetries and Reduction

The Hamiltonian formulation of analytical mechanics puts the dynamical equa-

tions of motion of conservative mechanical systems into the following form (known
as Hamiltonian equations) :

dH
4, = & (41,. . >qn,p,, . ,P,)
I
i= I,...,n, (1)
dH
P! = - j$q, ,...1 qn,p, ,...,p,,)

where q,, i = 1,. . . , n, are the generalized configuration variables, p,, i = 1,. . . , n,
are the generalized momentum variables and H(q,p) is the Hamiltonian function
which usually can be equated with the total energy of the system. The Hamiltonian
equations of motion can be derived from and are equivalent to the Euler-Lagrange
equations of motion (derived from a variational point of view, i.e. Hamilton’s
principle which in turn rests upon d’Alembert’s principle of virtual work) in case
the Hessian matrix of H(q,p) with respect to p is invertible (the so-called hyper-
regular case which is mostly satisfied). The generalized momentum variables pZcan
be defined through the generalized velocities in the Lagrangian function ; however
the real power of the Hamiltonian formalism (which is at the heart of theoretical
physics) lies in treating the variables q and p at the same level.
An important mathematical achievement of the last century has been the coor-
dinute ,free definition of the Hamiltonian equations (1) in terms of symplectic
manifolds (34,35) ; for a complete modern account of this topic see (I, 2). Here
ql,.‘.,q,,~Pl,..., p,,) are seen as speciul local coordinates for a manifold A4 (the
phase space) which is endowed with a particular geometric structure called “sym-
plectic structure” (defined by a non-degenerate closed two-form o on M). In fact,
these coordinates are adapted to w in the sense that COtakes a particularly simple
form in these coordinates : it can be equated with the constant bilinear form defined
by the matrix :

G-9

The special coordinates (q,, , y,,, p,, .p,) are called “canonical coordi-
nates”. Subsequently it was realized that an equivalent geometric definition of the
Hamiltonian equations can be given in terms of a non-degenerate Poisson structure
(which mathematically is dual to a symplectic structure (5)). The expression in local
coordinates of such a non-degenerate Poisson structure is nothing other than
the classical Poisson bracket. The geometric coordinate-free formulation of the
Hamiltonian equations (1) is now performed by identifying a phase space M
endowed with a symplectic form (1)(or equivalently with a non-degenerate Poisson
structure), together with a Hamiltonian function H: M + R. The triplet (M, CO,H)
determines the dynamics (1) uniquely by the fact that the symplectic structure
defines a mapping from functions on M, in particular from H, to vector fields on
M.

Vol. 329. No. 5, pp. 9235966, 1992

Pmlkd I” Great Bntam 925
B. M. Muschke et al.

Recently it was recognized that an important generalization of the standard

Hamiltonian equations (1) can be obtained by not being restricted to a non-
degenerate Poisson structure, but instead to allow for general ones. General Poisson
structures have been already introduced by Lie (28) mainly for studying systems
of partial differential equations of first order, e.g. Hamilton-Jacobi equations (see
also (29)) but were rediscovered only recently (3&32). The relevance of such
general Poisson structures has already become evident in the study of reduced-
order systems arising from systems having symmetries, such as the Euler equations
for the motion of the rigid body or more complex systems with symmetries (537).
The basic connection of the Euler equations with the canonical Poisson structure
on the dual of the space of 3 x 3 skew-symmetric matrices (see Examples 3 and 7)
was identified in (38). Also, for infinite dimensional systems the Poisson structure
point of view has proven to be very valuable, see e.g. (35).
A main purpose of this paper is to show that Poisson structures appear naturally
in the dynamic equations associated with a network representation of physical
systems (here restricted to systems without dissipative phenomena).

2.1. Poisson structures and Humiltonian systems

We will now briefly recapitulate the basic mathematical ingredients of general
Poisson structures and the resulting generalized Hamiltonian equations of motion
as alluded to in the introduction to this section. Because of space limitations we
can only sketch the outline of the theory. For a more elaborate treatment we refer
the reader to (5,30,35), where detailed proofs can also be found.
We start with the definition of general Poisson structures, which constitutes the
basis of the definition of generalized Hamiltonian systems.

Let M be a smooth (i.e. Cm) manifold and let Cm(M) denote the smooth real
functions on M. A Poisson structure on M is a bilinear map from C-(M) x Cm(M)
into C”(M), called the Poisson bracket and denoted as :

(F,G)H{F,G}ECI‘(M), F,GEC~(M),

which satisfies for every F, G, HE C”(M) the following properties :

skew-symmetry :

{F,G} = -{G,FS, (3)

Jacobi identity :

IF,{G,H}}+{G,{H,Ff}+{H,CF,G}} =O, (4)

Leibniz rule :

{F,G-H} = {F,G}-HfG-{F,Hj. (3

M together with the Poisson structure is called a Poisson manifold.

Now let A4 be a Poisson manifold with Poisson bracket { , }. Then for any
HE C”(M) and arbitrary x E M, we can define the mapping :

Journal of the Franklm Inst~luw

926 Pergamon Press Ltd
 Humiltonian Formulation of Network Dynamics

X,(x) : C”(M) -+ R

as

x,(x)(F) = {F, H}(x), FE C” (MI. (6)

It follows from the bilinearity of the Poisson bracket and (5) that X,(x) E T,M for
every XEM (where T,M denotes the tangent space to M at the point XE M).
Consequently X, defines a smooth vector field on M, called the Hamiltonian vector
fieldcorresponding to the Hamiltonian function H (and with respect to the Poisson
bracket { , }). Thus (6) can be seen as the (Lie) derivative of the function F along
the vector field X,.
Note that this definition implies that the Hamiltonian His necessarily a conserved
quantity for the Hamiltonian vector field X,. Indeed by using skew-symmetry (3)
we have by (6) :

X,,(x)(H) = {H,Hj(x) = 0. (7)

Furthermore, we note that since

{F> G) = x,(F)(x) = dF(x)(&(x)),

{F, G} = -(G, F}(x) = -X,(G)(x) = -dG(x)(X,(x)) (8)
the value of the Poisson bracket {F, G} in any point x E M only depends on the
differentials of F and G in that same point. Therefore the mapping HE C”(M) to
X,(x) E T,M for any x E M, as defined by (6) can be also seen as a mapping from
dH(x) E T,*M (the cotangent space to M at XEM) to X,(x) E T,M. Hence the
Poisson structure defines for any x E M a (linear) mapping from CM to ll”,M.
A more concrete representation of this mapping is obtained by taking local
coordinates x,, . . . ,x, for M (defined on a neighborhood of some point x,, E M).
Since the Poisson bracket {F, G}(x) only depends on the differentials of F and G
in x, we obtain the following coordinate expression for the Poisson bracket :

for some functions Jk,~ C”(M), k, I = 1,. . . , m, which in fact are determined by :

%,(_x) = {x~,x,}(x), k,l= l,..., m. (10)

(Here xk,xj are seen as coordinate functions on M, i.e. as elements of C”(M).)
From (3) we obtain

Jk,(_x) = - JIk(x), k, I = 1,. . . . , m (11)

while (4) yields

J,,$(x)+J,~(x)~(x)+J,~~ (x) = 0, i,j,k = I,..., m. (12)

I 1 I

Conversely, if the smooth functions JII, k, I = 1, . . , m, defined locally on M, satisfy

(11) and (12) then one can define locally the Poisson bracket { , } as in (9) and it

Vol. 329, No. 5, pp. 923-966, 1992

Prmted in Great Bntain 927
B. M. Muschke et al.

may be verified that the Poisson bracket verifies (3) (4) (5). In conclusion, locally
the Poisson bracket is uniquely determined by its structure mrtris :

J(x) = (Jd.~)h.,= I. _,,il (13)

with JA, satisfying (I I), (12). Furthermore it follows from (6), that the Hamiltonian
vector field X, expressed in local coordinates x,, . ._Y,,,as the vector (X:,, . . ,X;‘;)’
is given as :

.
. J(s) ...

I .
(14)
1/q;(s) (y (x)
111

and thus the dynamical equations of motion determined by XH read in local

coordinates as :

= J(s) (15)

The above also implies that the map :

J(x) : TfMw T, M (16)

can be seen as the local matrix representation of a bundle map from the cotangent
bundle T*M to the tangent bundle TM.
Let us remark at this point that for the definition of XH(x)~ T,M in (6) or the
coordinate representation J(x) : T,*M t-+ T,M in (I 3), we did not yet use the Jacobi
identity (4) or equivalently (I 2) ; i.e. we could have defined everything without this
assumption on the bracket ( , ). The Jacobi identity does however play a crucial
role in the following property of any Hamiltonian vector field X,,. Let F, GE CL (M),
then by (6) the Jacobi identity (4) and (3) :

X,({F,G})= (H,{F,G)) = { N,F),G +{F,{H,G}) (17)

and thus :

x,,({C G)) = Ix,(F), G) + {F,XdG)). (18)

Property (18) means that X, leaz~~s the Poisson bracket itmwiant. Indeed, if WC
denote by XL the time t integral of X,, (i.e. Xj,(x,) is the solution at time t of
.e = X,(.\-) with -u(O) = .Y”) then (18) is equivalent to :
 Hamiltonian Formulation of Network Dynamics

{FOX;,GOX;}(~) = {F,G)(x) VF,G~EC~((M), Vt 3 0. (19)

Conversely, given a bracket { , }, satisfying (3) and (5) one can define for each real
valued function H on R, a vector field _I’, as in (6). Then the bracket satisfies the
Jacobi identity (4) (and thus is Poisson bracket) if and only if for each H the
Hamiltonian vector field X, leaves the bracket { , > invariant (i.e. satisfies (18)
(19)).
The rank of the Poisson bracket { , } in any point XEM, is defined as the rank
of the structure matrix J(x) in this point. (This can be shown to be independent of
the choice of the local coordinates.) By (11) necessarily the rank at any point is
even. A Poisson manifold A4 having the property that the rank of the Poisson
bracket is equal everywhere to the dimension of A4 is called a sympfectic manifold.
(Thus necessarily the dimension of a symplectic manifold is even.) On the other
hand a general Poisson manifold M can be seen as a union of symplectic manifolds
which fit together in a smooth way in the following sense: through each point of M
there passes a unique submanifold with the property that the rank of the Poisson
bracket in every point of this submanifold is equal to the dimension of the sub-
manifold (5,30). In fact (see (5)) this submanifold is the maximal integral manifold
(through x0) of the distribution :

span {X,-(x)IFEC?(M)}, x~A4. (20)

Example 1. (Standard Hamiltonian equations.) Let A4 = R2” with natural coor-

dinates (x,, . , xzn) = (q ,, , qn,p ,, . . ,p,) and define the standard Poisson
bracket :

(21)

Notice that

J(x) = x E R2”,

and thus the rank equals 2n everywhere. For any Hamiltonian HEC” (Rzn), the
Hamiltonian vector field X, is given by the familiar equations of motion (c$ (15)) :

4, = ““(4>P)> d! = - “‘1(4,P), i= l,...,n (22)

8P; aqi

called the standard Hamiltonian equations. M = R*” with bracket (21) is a sym-
plectic manifold and q = (q,, . . , qn) andp = (p,, . . ,P,~) are called the generalized
configuration coordinates, respectively generalized momenta. Using the (inverse)
Legendre transform :

H(q,p) - Uq, 4) = i P! 4t - H(q, p)

i* I

Vol. 329. No. 5. pp. 923-966, 1992

Primed m Great Britain 929
B. M. Musckke et al.

with p determined by 4 = c?H/t?pone may arrive at the equivalent Euler-Lagrange

equations.

Exumple 2. Consider a simple linear electrical circuit consisting of two

inductances L,, L2 and one capacitor C (see Fig. I).
The total energy stored in the three elements is given as

(23)

where 4,, 42 denote the fluxes of inductances L, and L?, respectively, and Q is the
charge of the capacitor C. Using the constitutive relations of L,, L2 and C together
with Kirchhoff’s laws, we obtain the equation of motion :

(24)

which can be rewritten as :

This defines a Hamiltonian vector field on M = R7 corresponding to the

Hamiltonian H(Q, 4 ,, c$J and the Poisson bracket on IR’ given in natural coor-
dinates for Iw’ by the constant structure matrix :

(26)

Notice that J is independent from the parameters L ,, Lz and C, and in fact is solely

il

FIG. I. An electrical circuit

930
 Hamiltonian Formulution of Network Dynamics

determined by the topological structure of the network. In particular, (25) remains

valid if the elements are nonlinear and thus define a non-quadratic Hamiltonian
H(Q,@,,c$~). Clearly the rank of J is everywhere equal to 2, and M = R’ is the
union of all the two-dimensional symplectic manifolds :

{(Q>41~&)14, +4z = kJ> (27)

with k denoting a constant.

Example 3. (The rigid body.) Consider [w’ with natural coordinates denoted as
(p,,p,,p2). It can be verified that the matrix :

J(p) = [_(lp,?; -{ -;:j, P = (P,,?P!?P;)‘ER3 (28)

satisfies (11) and (12) and thus is the structure matrix of a Poisson structure on M.
Consider furthermore the Hamiltonian function :

(29)

Then by (14) and (15) we obtain the following equations of motion :

which are the well-known Euler equutions for the motion of a rigid body spinning
around its centre of mass (pI,pv,pc are the body angular momenta, see (35)). Note
that rank J(p) = 2 at every point (p.,,p,.,p,) # (O,O,O), and that M = R3 can be
seen as the union of the symplectic submanifolds of the form :

{(P,.,P,.,P,)~P,+P,‘+P_’ = kJ (31)

with K being a constant

The Poisson bracket introduced in Example 3 is an example of a much more

general construction. Indeed, let V be a matrix Lie algebra, i.e. a linear space of
n x n matrices which is closed under the matrix commutator

[A, B] = AB- BA. (32)

Then the dual space V* has a natural Poisson structure, called Lie-Poisson struc-
ture. Indeed, let F and G be real functions on V*. Then the differentials dF(x),
dG(x), XE V* (i.e. the row vectors of partial derivatives) are elements of (V*)*,
and thus can be identified with elements in V. Using this identification, one defines

(6 G}(x) = <-u, [dF(x), Wx)l), XE V*, (33)

where ( , ) denotes the pairing between V and V* and [ , ] is the matrix commutator
(32). It can be verified that { ,} satisfies (3), (4) and (5) and thus defines a Poisson

Vol. 329. No. 5, pp. 923-966. 1992

Printed in Great Britain 931
B. M. Maschke et al.

bracket on I’*. In more concrete terms, let z’,, . . , c, be a basis of V. Then C’,EV
can be identified with an element in (I’*)*, i.e. with a linear coordinate function
,yion V*,i= l,..., m.Nowdefine

Jli(X) = {.X,,.Y,} = [I’,, [‘,I, XE v*, (34)

then J(x) = [J,iWlz.,= I.. .mis the structure matrix of the Poisson bracket defined in
(33) and satisfies (11) and (12).
Example 3 fits into this theory as follows: the matrix Lie group underlying to
rigid body motion is SO(3), the space of 3 x 3 orthonormal matrices with matrix
algebra SO(~), the space of skew-symmetric 3 x 3 matrices. Thus let us consider
V = SO(~), with the standard basis :

and commutation relations :

[t.,.Z’J = 1’3, [@2,2’3]= I’,, [L’j,U,] = 2’2. (36)

Using (34) one sees that the Lie-Poisson bracket on V* = SO*(~) is given by the
structure matrix :

(37)

and thus by identifying M in Example 3 with SO*(~), i.e. identifying (p,,p, ,p_)‘with
(s ,, x2, xJ, one recovers the structure matrix (28).

2.2. S~~mmetries uncl Humiltoniun reduction

We start with the following basic theorem on general Poisson brackets (essen-
tially due to Lie (28)) which generalizes the definition of canonical coordinates
(q ,, . . , q,,,p ,, . . ,p,) for symplectic manifolds to general Poisson manifolds.

Theorem I
Let A4 be an m-dimensional Poisson manifold. Suppose the Poisson bracket has
constant rank 2n in a neighborhood of a point x,) E M. Then locally around x,, we
canfindcoordinates(q,p,r)=(q, ,..., q,,,p, ,..., p,,.r ,,..., r,)forMofdimension
(2~ + I) satisfying :

(%,P,) = 4,, (4,3Y,) = {Pl,P,) = 0, i,j = 1>..., n

f(q,, rl,} = (p, , r,, I_=
{r,~
,
r, I’ -0.
- i= l,...,n, i’,,j’= l...., 1 (38)

or equivalently, the m x m structure matrix J(q,p, r) is given as :

 Hamiltonian Formulation of Network DJ3namic.s

(39)

The coordinates (q,p, r) satisfying (38) are called (generalized) canonical coor-
dinates.

Remark 1. If M is a symplectic manifold then I= 0, m = 2n and r is void, thus

leading to the structure matrix :

(40)

Hence Theorem I implies in particular that on every symplectic manifold we can

take local coordinates q,, . . , qn,p,, . . . ,pn in which the Poisson bracket takes the
form of the standurd Poisson bracket, see Example 1.

Let M be an m-dimensional Poisson manifold with Poisson bracket of constant

rank 2n in a neighborhood of a point SUE M. Then in view of (14) and (15) in
the local canonical coordinates of Theorem I any Hamiltonian vector field X,,
H: M H Iw, takes the form

fj, = $(q,p,r),
I /
i;=0,
li, = - i:(q,p,r). i= l,..., n, ,j= I,..., I,

(41)
which generalizes the standard Hamiltonian equations (1).

Example 4. Let M be a symplectic manifold with Hamiltonian vector field X,,,

HE C”(M). Then by Remark 1 and (41) one can find local canonical coordinates
ql,...,q,,pl,.“> pn such that the equations of motion corresponding to X, are
locally given by the standard Hamiltonian equations (22).

For any Poisson manifold we define the distinguished or Casimir functions as those
smooth functions F: MH IL!,such that :

(F,G} = 0, VGEP(M) (42)

or equivalently, since {F, G} = -X,;(G) (see (6)): A’,..= 0. Hence the Casimir
functions correspond to the kernel of the map : Ft+ X,.. from C”(M) modulo 178to
the vector fields on M given by (6). In local terms F is a Casimir function if
dF(x)E kerJ(x), i.e. if dF(x) is in the kernel of the map J(x): T,fMt+ T,M foi
every x. Under the assumptions of Theorem I, the Casimir functions are locally
given by the coordinate functions r,, . . . , r, together with all functions on Mdepend-
ing only on r,, , r,.
Theorem I has some interesting consequences concerning the local embedding
of a Poisson manifold into a higher-dimensional symplectic manifold, as well as
the local reduction of a Poisson manifold to a lower-dimensional symplectic,

Vol. 329, No. 5, pp 923-966. 1992

Printed in Great Britam 933
B. M. Muschke et al.

manifold. First, for any two Poisson manifolds M, and Mz, one defines a smooth
map : C#I: M, H M2 to be a Poisson (or canonical) mapping if:

{P~,Gc 4;, = {F,G}~‘J~,~F,GEC’(M~) (43)

where(,)ra , >>denote the Poisson brackets on M, and M?, respectively. Now

consider the local coordinates (q,p, Y) of Theorem 1. Then the map rr : R’“+‘H R2”
given by :

(q,,...,q,,p,,...,p,,,r,,...,r,);t(q,,...,q,,p,,...,p,,) (44)

is a Poisson mupping from the Euclidean space RZn+’ with Poisson bracket given
by (39) to the lower dimensional space R2” endowed with the standard Poisson
bracket (40). On the other hand, consider the extended space R”‘+” with natural
coordinates (q,, . . . , qn,pI,. . . ,pn,sI,. . . .s,, Y,, . . . , r,) and standard Poisson
bracket given by (38) together with the additional relations :

(s,,r,> = 6,, , i’..j’ = 1, . . , I

(ql,“,,}={pi,s,,}=O, i=l,..., n, i’,j’= l,..., 1. (45)

Then the map 7c: Rz”+2’t+ tR’“+’ given by :

is also a Poisson mapping. Hence in view of Theorem I, we have the following

picture in a neighborhood of any point x0 in the (2n+I)-dimensional Poisson
manifold M where the Poisson bracket has constant rank 2n :

u” c [W2n+2/c u c [w2n+/;, u’ c RZ”,

(47)

where U, U“, U’ are open parts of R2”+‘, R2”+“, R2”, respectively. Hence the
coordinate neighborhood U can be reduced (via 7~)to a symplectic space (i’ and
can be embedded (via FL) into a symplectic space U’. (Furthermore it can be shown
that II and rc are unique in the sense that if (47) holds for difSerent Poisson mappings
n’ and II’ then there exist Poisson mappings $ : RZn++ RZn, ‘P : R2”+2’~ lR2”+2’ of
maximal rank such that Y rc’ = n, ‘P 1 TI’ = FI (see (30).)
The above discussion on local normal forms of Poisson structures with constant
rank has the following implications for the structure of any Hamiltonian vector
field on the Poisson manifold M in the neighborhood of a point x0 E M where the
Poisson bracket has constant rank. Let H: MH R be a Hamiltonian function, then
locally about x0 the Hamiltonian vector field X, is given in the local coordinates
(q,p, r) of Theorem I by (41) i.e.

G,=e(q,p,r), p,= _dH

aq,(q~p~rh i,=O, i=l,..., n, j= l,..., 1.
dPi
(48)

934
 Uamiltonian Formulation qfNetwork Dynamics

Following (47) these equations can be restricted for every constant (r!, . . , r,) to
the standard Hamiltonian equations on R2” (c$ (1) or (22))

(49)

where W: iR2n~ R is given by W(q,p) = H(q,p, r), and can be embedded into the
standard Hamiltonian equations on IL!‘“+“:

i= 1,. ..,n, j= l,..., 1. (50)

Note that in (50) the Hamiltonian H seen as a function on iR*‘+ 2’does not depend
on s, while in (49) the Hamiltonian I-p(q,p) = H(q,p, r) is seen as a function on
D%‘” parametrized by Y.
The transition from the standard Hamiltonian equations (SO) on IL!“+” into the
standard Hamiltonian equations on (49) on R2n, via the non-standard Hamiltonian
equations (48) can be interpreted as a canonical reduction of order caused by
symmetry of the HamiItonian function, which is a vast subject in analytical mech-
anics (see e.g. (2,39-41)). Indeed the crucial point is that Hin (50) does not depend
on the canonical coordinates s,, _. . ,s,, or equivalently that H is invariant under
the infinitesimal canonical mappings a/&,, . . , a/&,. It is well known in the theory
of Hamiltonian systems with symmetry (see (35)), that this implies that RZ’+2’may
be projected to IR‘“+‘with Poisson structure given by (38), and that the Hamiltonian
dynamics (50) project to (48). Furthermore the Casimir functions r I, . . . , rl for (48)
are conserved quantities and thus (48) can be restricted (for any constant r,, . . . , vi)
to (49). Thus alternatively the Poisson manifold M, locally given as iR2nC’ with
Poisson structure (38) and the Hamiltonian dynamics (41) defined on it, may be
considered as an intermediate step of the canonical reduction of order of the
standard Hamiltonian equations (50) to the standard HamiItonian equations (49).

Example 5. Consider the electrical circuit of Example 2, with Poisson structure

determined by (26). Canonical coordinates are given by

4=e7 P=:(42-#1), r=4,-42 (51)

and the total energy (23) expressed in these coordinates is

H(q,p, r) = g +
(52)

resulting in the dynamical equations (cz (52)) :

r
--P ;+P
g=iz(q,p,r)= -Q+-z-, p= -g(q,p,r)= Q
.-?+ r: = 0. (53)
L

Vol. 329, No. 5. pp. 923-966.1992

Printed in Great Britain 935
5. M. Maschke et al.

The last equation reflects the fact that r = d,, +c$? (the total flux) is a Casimir
function, and thus a conserved quantity.

Esarrzplc 6. Consider the mass-spring system described in Fig. 2 which constitutes

the mechanical analog of the electrical circuit described in Example 2. Its total
energy is given by

(54)

with p,, p1 denoting the momenta of mass IJI, and w?, respectively, and s,’ the
distance between them and k denoting the spring constant. The Poisson structure
is again given by (23). resulting in the equations of motion (compare with (22)) :

(55)

Canonical coordinates are given, analogously to (5 I), as

(/ = .X,2, I’ = i(p2-PI), I’= PI +t1)2. (56)

where the Casimir function r = p, +p2 is the total momentum of the mechanical
system. The system can be embedded into standard Hamiltonian equations on R“
(r:fI (50) for n = 2) by adding a variable s satisfying (c$ (45)) :

(s,rj = I, ;q,.Y) = {p,s) = 0. (57)

The resulting standard equations are

I I I

iiH 2+p
li=-kg,S=~clr(‘I,P,“)=- + i=O
ng I m2 ’

which is isomorphic, using the canonical mapping implicitly defined by (56) and
s = J(.u, +X2), X,2 = x,--x,, to

FIG. 2. A one-dimensional mechanical system.

936
 Hamiltonian Formulation of Network Dynamics

Pi P2
I, =-) it,=--, Li, = W-x,), b2 = -k(x2-x,), (59)
mI m2

which are the standard Hamiltonian equations of motion for the mechanical
system, where x, and x2 denote the configuration coordinates of masses m, and
m,, respectively, in an inertial frame.

Example 7. Consider the Euler equations for the motion of a rigid body as dealt
with in Example 3, withp,,p,,pz the body angular momentum variables. Canonical
coordinates (outside the origin) are, for instance, (see e.g. (35)) :

q =p_, p = arctan&, r = p2 +p: +pl.

PX
Notice that r is the total angular momentum, which indeed is a conserved quantity.
The embedding standard Hamiltonian system (50) in this case does not seem to
have an easy physical interpretation, and will not be specified. (Notice that in (39),
example 3.2) it is briefly indicated how alternatively a 4-dimensional embedding
Hamiltonian system may be obtained by using the fact that SO*(~) is isomorphic
to the complex matrix Lie algebra su(2), which acts symplectically on the symplectic
space @*.)
Of course the rank of the Lie-Poisson structure is zero at the origin. Thus a truly
globally defined embedding Hamiltonian system has to be 6-dimensional and, in
fact, may be obtained in the following way (see (2) for details). The configuration
space for the dynamics of a rigid body around its centre of mass is SO(3), the
matrix group of 3 x 3 orthogonal matrices. Hence the phase space of the standard
Hamiltonian equations for the rigid body equals T*SO(3). Now T*SO(3) can be
identified with SO(3) x SO*(~) in the following way (see e.g. (2)). Define for any
gE SO(3) the left translation L, : SO(3) H SO(3) as :

L,qh = gh, h E SO(3). (61)

Denoting the 3 x 3 identity matrix Z3E SO(3) by e, we obtain the diffeomorphism :

L,* : Z-,*SO(3) H T,*So(3) = : So*(3) (62)

and the resulting diffeomorphism A: T,*SO(3) F+ SO(3) x SO*(~) defined as

A@, lx) = (g, L,*a), gE SO(3), a E T,*SO(3). (63)

Identifying in this way T*S0(3) with SO(3) x SO*(~) we see that the Hamiltonian
function H: SO(3) x SO*(~)++ R, being the kinetic energy, only depends on the
natural coordinates pX, pY, p_ for SO*(~), and in fact is given as in Example 3. Hence
the embedding standard Hamiltonian equations are in local coordinates of the
form (50) and thus project to the Euler equations (30) on the Poisson manifold
SO(3) x so*(3)/So(3) = So*(3).
Given the standard Hamiltonian equations (50) where H does not depend on
the canonical variables s = (s,, . , s,), there exists also an alternative reduction

Vol. 329. No. 5, pp. 923-966. 1992

Printed m Great Britain 937
B. M. Maschke et al.

procedure which leads eventually to the same reduced standard Hamiltonian system
(49) but not via the non-standard Hamiltonian equations (48). Instead, one uses
the fact that r = (r,, . , r,) are ~orlspn~edquantities for (50) from which it follows
that (50) can be restricted for every constant vector r to dynamics R*“+‘, and
subsequently may be projected to the same standard Hamiltonian dynamics (48).
Pictorially both reduction schemes can be summarized by the commutative
Diag. 1.
In the present context of network representation of physical systems, however,
the left-hand side reduction procedure appears to be the most natural one. Since this
reduction procedure starts with the invariance (or symmetry) of the Hamiltonian
(internal energy), the conserved quantities become fully captured in the Poisson
structure, i.e. the ycorn~tr~~ underlying the general Hamiltonian equations (48).
Indeed the conserved quantities are determined by the Casimir functions for this
Poisson structure, irxl~qendec~tl~ from the Hamiltonian of (48).

Remark 2. Let us mention that we have restricted ourselves to a purely loc.ul

description of the reduction procedure from equations (50) to (49), via (48). For a
global and coordinate free treatment, we refer to e.g. (35). see also (2). Ofparticular
interest is the case of general Hamiltonian dynamics (48) defined on a Poisson
manifold which is the dual of some Lie algebra V (endowed with the LieePoisson
bracket. c;f: the discussion after example 3) such as SO*(~). Then an embedding
standard Hamiltonian system is to be found by looking for a symplectic action of
the corresponding (simply connected covering) Lie group on some symplectic
manifold M (see (30,41)). In this case the momentum map M -+ V* (see (2)) provides
the projection map from (50) to (48), and the embedding standard Hamiltonian
system is living on this manifold M.

Remark 3. In (41) it is shown how the dynamics of incompressible fluids,

described as general Hamiltonian equations of motion on the dual of the Lie
algebra corresponding to the group of volume-preserving diffeomorphisms, can be
embedded into a standard Hamiltonian system with canonical coordinates given
by the classical Clebsch variables.

proJectIon X restriction
H( q. y. I; s )

L/ of eq. (50) \
R (21+/J
R (al+!)

of eq. (49)

DIAG. I

938
 Hamiltonian Formulation qf Network Dynamics

III. Bond Graph Models and their Dynamics

3.1. Physical systems in terms of bondgraphs

The modern graphical description of physical systems emerged from the descrip-
tion of electrical circuits. Kirchhoff introduced, from the inspection of its material
organization, the first graphical type abstraction into constitutive relations associ-
ated with electrical elements (R, L, C) and the topology defining the circuit (Kirch-
hoff’s current and voltage laws). Although mechanical systems provided the
framework for a general analytical formalism of physical systems models, leading
to analytical mechanics, it was also soon recognized that circuit theory is not
restricted to electrical circuits. Indeed Maxwell already stated Newton’s second
law in a circuit-theoretical way, in the form of a balance equation.
The network formalism as a unifying conceptual framework for physical systems
was developed mainly in engineering (for historical developments see (21)) where
it revealed an extraordinary efficiency not only for the design of electrical circuits,
but also for industrial devices pertaining to different physical domains. For the
latter systems a particular graphical notation, called bond graphs, was developed
originally by Paynter (11). Its network structure called “junction structure”displays
the invariants of the physical system according to Tellegen’s and Kirchhoff’s
theorems on networks, but corresponds to an additional abstraction level. Indeed
it represents abstract energy flows between elements and no more material inter-
connections. In the same way, the bond graph formalism is based on a classification
of physical variables, rooted in thermodynamics (2325). This led to the generalized
bond graph formalism (13) based on an abstraction of the elements of a network
(their phenomenological properties) and the systematic use of a unit gyrator (the
gyrator element was invented by Tellegen). This element, called “symplectic gyra-
tor” (25) will be a fundamental link relating analytical mechanics with the network
formalism.

3.1 .l. Energy and the capacitor element. The elemental components of a bond
graph model are elemental systems endowed with energy, called (multiport) energy
storage elements and denoted by C (11, chap. 4; 13, chap. 5; 25).
Such an elemental system concentrates some physical properties inside abstract
boundaries defining it with respect to its environment. Its internal state is defined
by a vector x in [w”of energy variables and a real-valued smooth energy function
H(x) describing its static characteristics. The existence of an energy-function cor-
responds physically to the first law of thermodynamics. Then the variation of the
energy and the energy variables are related by Gibbs fundamental equation :

dH = ,$, g dx, . (64)

The interactions with the environment through the boundaries of the elemental
system are defined by the time-variation of the energy at the ports of the elemental
systems. The time-variation of the energy may be expressed by two vectors of so-
called conjugated power variables called “effort” e and “flow” f:

Vol. 329, No 5, pp 923-966, 1992

Prmed m Great Britain 939
B. M. Maschke et al.

f= i = (i,, . . ,a,)‘, e=dH=(g,...,iE). (65)

The variation of the energy function is then the inner product of these two vectors :

dH n
dt = i;,el*.f;. (66)

It may be noticed that as a consequence of the existence of an energy function, the

effort as a function of the energy variables verifies the Maxwell reciprocity relations :

de, de.
1 i,jEl,...,?I (67)
a41 a4i ’
(and by Poincare’s lemma the converse also holds).
In summary the energy storage elements C are represented in bond graph terms
as in Fig. 3.

E.xample 8. (Point-mass.) A point-mass may be represented as a l-port energy

storage element (in the kinetic domain) with the translational momentum p as
energy variable and the kinetic energy Hki”(p) = p2/2m as energy function, where
m denotes the mass. The power variables at its port are the forces acting on the
mass (i.e. the rate of change of the momentum) as flow variable and the translational
velocity of the point-mass as effort variable.

E,xample 9. (Electrical capacitor.) An electrical capacitor may also be represented

as a l-port energy storage element with the electrical charge q as energy variable
and the electrical energy as energy function (in the linear case: q2/2C, where C
denotes the linear capacitance). The flow variable is the current through the capaci-
tor (i.e. the rate of change of the charge) and the effort variable is the voltage at
its port.

Example 10. (Simple thermodynamical system.) A simple thermodynamical

system is characterized by an internal energy function U(S, V,N) (13, chap. 5).
Such a system may be represented as a 3-port energy storage element with the
number of moles N, the volume I/ and the entropy S as energy variables and the

C3H
e = k&+l..,t

H(x):a: t_1L ’
f=i
FIG. 3. An energy storage element

Journal oithc Franklin lnst~~tc

940 Pergamon Press Ltd
 Hamiltonian Formulation qfNetwork Dynamics

FIG. 4. A simple thermodynamical system.

internal energy as energy function (see Fig. 4). The conjugated effort variables are
the material potential ,U = aU/aN, the pressure P = -(~3UjdV’) and the absolute
temperature T = aUji3,S (where p is depending on P and T due to the first-order
homogeneous nature of the energy U(S, V, N) (13, chap. 5)). Such a 3-port energy
storage element can be used, for instance, to model an open container with piston
and containing a gas.

3.1.2. Dynamical interactions unci the sJ?mplectic gyrator. In the previous section
we have defined elemental systems endowed with energy, but which cannot undergo
any dynamic evolution per se. Any dynamics arises from the interaction between
elemental systems. These interactions are described as relations between the effort
and flow variables, called phenomenological laws like Newton’s or Ohm’s laws. In
bond graph terms these interactions are described by elements whose constitutive
relation represents a phenomenological or constitutive law and are power continu-
ous, i.e. the total power flow at their ports is equal to zero (as any energy storage,
i.e. any energy function, should be associated with an energy storage element C).
There are two types of such elements.
The resistive irreversible transducer relates some physical domain with the ther-
mal domain and describes resistive laws (lS15). This element will not be considered
in the present paper, as we have restricted ourselves to the so-called “conservative”
systems without any resistive or, rather, irreversible phenomena. The other element
relates two different physical domains : either the kinetic with the elastic (potential)
domain, or the electrical with the magnetic domain. The kinetic and elastic domains
include the special forms which are present in hydraulics or acoustics. This second
element is called “symplectic gyrator” (25). It is a 2n-port element whose consti-
tutive relation is, if,f and e denote the (2n-dimensional) flow and effort vectors at
its ports and PymP the symplectic matrix of order 2n :

where I, is the identity matrix of order n.

Example 11. (Mass-spring system.) Consider the mass-spring system and its
bond graph representation depicted in Fig. 5. With the spring a one-port energy
storage element is associated : it defines the elastic potential energy of the system
as a function H,,(q) of the displacement q (the energy variable of the elastic domain).
For a linear element the energy function is :

Vol. 329. No. 5. pp. 923-966. 1992

Prmed in Great Britain 941
B. M. Maschke et al.

-J spp
cik et, - u
!t&,C~()~C2 0-a: g
ef- d
-I;.. x,.-i
(b)
FIG. 5. Mass-spring interaction and bond graph model.

H,,(q) = 4kq2, (69)

with k the stiffness of the spring. The point mass is represented by an energy storage
element with the momentum p of the mass as energy variable and the kinetic energy
Hkln@), which is in the linear case :

Hkm(P) = ; ; (70)

with nz the mass of the point mass. The complete bond graph model is obtained
by relating the two energy storage elements via so-called bonds through a symplectic
gyrator. The bonds may be considered as identities on the power variables, i.e.
efforts and flow variables of the related ports of the elements. (The next section
will give the proper definition of the bonds). The symplectic gyrator represents the
following two relations. First the attachment of the end of the spring to the mass
in globally Eulerian but locally Lagrangian coordinates with respect to an inertial
frame (13, chap. 6) :

(71)

The variable ekln is the effort in the kinetic domain, i.e. the velocity of the mass I’,,,~,~
and ,fC,is the flow of the electric-potential domain. The second relation represents
Newton’s second law in an inertial reference frame :

g = .fkin = -e,, = -F,, . (72)

The variable e,, is the effort in the elastic potential domain, i.e. the force induced
by the spring, and ,fkin is the flow of the kinetic domain, i.e. the time variation of
the momentum of the mass.

942
 Hamiltonian Formulation qf Network Dynamics

Example 12. (Inductor and capacitance circuit.) Consider the electrical circuit
of Fig. 6 and its bond graph model. The electrical domain is represented by an
energy storage element C with energy variable Q (the electrical charge of the
capacitor) and the energy function H,,(Q) which is, for instance, with the assump-
tion of the linearity of the constitutive relation :

H el(Q) = ’ c (73)
2c

with C the capacitance.

The magnetic domain is represented by an energy storage element C, with the
magnetic flux linkage 4 as energy variable and energy function H,,,,(4) which is,
for instance, with the assumption of linearity of the constitutive relation :

with L the inductance.

The symplectic gyrator represents Maxwell’s second and fourth law coupling the
electric and magnetic fields in the quasi-stationary approximation of electrical
circuits (25) according to

dQ
~=LI =i=emg, ;; =J;n&, = -_u = -e,,. (75)

The generalized bond graph formalism recalled here makes a gyrative coupling
between different physical domains appear explicitly, which would not appear in
the “conventional” bond graph approach (as well as in electrical networks). In the

(4

(b)
FIG. 6. Inductor capacitance interaction and its bond graph model.

Vol. 329, No. 5, pp. 923-966, 1992

Pnnted m Great Britain 943
B. M. Mmchke et al.

g:c
“=I &_”
@2
_-'l_k~
.
i=y L=i
L
FIG. 7. Conventional model of the mass -spring interaction.

conventional bond graph approach, for both examples. there would be two kinds
of storage elements in interaction as represented in Fig. 7. The bond still represents a
coupling through power exchange, but using an operational identification (through
measurement) of across-variables (like velocities or potentials) and through-vari-
ables (like forces or currents). The definition of effort and flow variables are in this
sense less general than in the generalized bond graph formalism and are linked to
the concrete network. The two kinds of storage elements I and C have a dual
definition with respect to the power variables. In generalized bond graph terms the
I element is the aggregation of a C clement with a symplectic gyrator: this aggre-
gation is called “dualization” of the c’ element (13,25). Indeed the generalized
bond graph formalism prefers to give an asymmetric definition to flows and efforts
and introduces explicitly the gyrative couplings, based on the fact that in the
thermal domain there is no “inertance” and that in the conventional approach
some phenomena may be overlooked (25).

3.1 .3. Ponw c~or~titi~rit~~ NS NII uhstrrrct m~twwli postulut~~. Complex autonomous
physical systems may now be defined as a set of interacting elemental systems (i.e.
energy storage elements denoted by C) exchanging energy in such a way that it is
conserved over the whole system. The postulate of power continuity sets forth that
one may define the conservation of energy in the form of an abstract topology of
power flows between elemental systems (11,22). This topology of power flows is
called “junction structure”. Consequently the edges of this network transport
power flows and are called “power bonds” or simply “bonds”. In order to represent
the interaction between subsystems. for instance elemental systems. they satisfy the
“power postulate” : there exist two dual variables, called power variables, such
that their pairing is equal to the power flow. The two variables are called a flow .f
(usually a vector in W) and effort CJ(usually also identified with a vector in KY
although it is a dual quantity to the flow ,/). In summary a power bond is associated
with three variables (flow ,f; effort P and instantaneous power P) which verify :

(r,.f’) = P. (76)

In Fig. 8, the half arrow on the power bond indicates the positive orientation for

FIG. 8. Power (multi-) bond with sign orientation.

944
 Hamiltonian Formulation of Network Dynamics

-+h+
e2
A.. f2

FIG. 9. A O-junction.
en_
h-1

the power P and the flow variable f.The nodes of the junction structure are
necessarily power continuous elements and will be presented in the next paragraph.

3.1.4. Invariants displayed by generalized junction structures. The elements of a

generalized junction structure extend the representation of invariants of the physical
system from the energy (represented as energy storage elements) to structural
invariants induced by structural constraints (the topology of an electrical circuit
or kinematic constraints) or by interdomain couplings. Thereby the generalization
from quasi-statics to dynamics may be realized.
Two basic elements are the simple junctions, denoted by 0 and 1. Using scattering
variables it can be proven that they are the only possible realization of port
symmetric, power continuous elements (11,42). They represent invariants on the
dual power variables efforts and flows (11, chap. 6; 22). A O-junction corresponds
to “effort potentiality”, i.e. defines a common effort variable to all bonds connected
to it (see Fig. 9). The symmetry of the junction and its power continuity imply the
following constitutive relations :

e, = ... = e,, i E, - J; = 0, (77)

where ci are equal to plus or minus one depending on the orientation of the power
bonds (half arrows in Fig. 9).
A 1-junction represents “flow continuity”, i.e. common flow variables to all
bonds connected to it (see Fig. 10). It has the following constitutive relations
(dual to that of a O-junction) :

f, = ... =fn, i$,h*ei=O, (78)

where gi are the sign functions depending on the orientation of the power bonds.
Simple junction structures are networks of power bonds and junctions which

FIG. 10. A l-junction.

Vol. 329, No. 5, pp. 923-966, 1992

Printed in Great Bntain 945
B. M. Maschke et al.

describe sign-weighted invariants (i.e. constraints on the power variables). Such

local invariants may arise directly from the material organization of the system,
abstracted as a directed graph, for instance in the case of electrical circuits or one-
dimensional mechanical systems. In the case of electrical circuits the invariants
generated by the directed graph are called Kirchhoff’s voltage and current laws.
For such invariants, arising from graphical interconnection constraints, several
systematic procedures exist of the realization of the equivalent simple junction
structures (4246).

E.awnpl~~ 13. (Simple junction structure of an electrical circuit.) Let us consider

the electrical circuit of Example 2 (see Fig. I). The underlying directed graph
(with the elements associated with the branches indicated in brackets) and the
corresponding simple junction structure are represented in Fig. I 1. The constitutive
relations of the two simple junction structures represent, in the case of electrical
circuits. Kirchhoff’s current and voltage laws with the dual definition of currents
as either flow or effort variables.

a
B

1 2

A
(4

(b)

T&
i3
r
1+
VBA

vBA
Cc)

FIG. I I. Directed graph underlying an electrical circuit and the two dual junction structures.
 Hamiltonian Formulation qf Network Dynamics

FIG. 12. A causally augmented bond.

Hence a simple junction structure is constructed out of junctions representing

local invariants (in the sense that they are represented by distinguished nodes of
a bond graph), but it defines itself global invariants on the efforts and flows at
its ports (40, 41, 43). They are graphically displayed as a causal augmentation
of the bonds at the ports of the junction structure : a causal stroke is added at one
end of the bond (see Fig. 12). A causal stroke directed towards (or outwards,
respectively) a simple junction structure indicates an independent effort variable
and dependent flow variable (or independent flow and dependent effort, respec-
tively). The consistency of the invariants at the ports and at a junction in a simple
junction structure is ensured by causality restrictions at the junction (see Fig. 13)
and for particular bond loop structures by some global causality restrictions (i.e.
restrictions on some part of the simple junction structure (40,41)). For simple
junction structures representing electrical circuits, the causal augmentation at the
ports corresponds to the choice of a spanning tree or cotree of the underlying
directed graph (41, 43,44).

Example 14. (Causal restriction in an electrical circuit.) Consider the electrical

circuit of Example 2 and its directed graph in Fig. 11. Then a consistent causal
augmentation and choice of the spanning tree (here a single branch) is shown in
Fig. 14.
One may note that simple junction structures allow one to represent very general
power continuity relations arising from topological constraints in different kinds
of physical domains. For instance, the complex topological constraints of planar

-+lo+
LL
e2 f2 en_
. .
h-1 n

+ 1 I+
e2
L-L
.. ”
fi en_l

FIG. 13. Causal restrictions

Ad
at junctions.

Vol. 329. No. 5. pp. 923-966, 1992

Printed in Great Britain 947
 B. M. Maschke et al.

1 2 3

@ A
(4

lb)
FK. 14. Choice of causality and associated spanning tree.

and spatial mechanical linkages described in skeleton diagrams or interchange

graphs may be represented in kinematic junction structures using generalized junc-
tions (47,48).
An additional power continuous element, called “transformer” and denoted by
TF, generalizes the sign-weighted relations expressed in simple junction structures
to linear relations. A junction structure containing transformers is called weighted
,junction structure. A multiport transformer imposes linear relations between the
power variables (f2, eJ of port 1 with dimension n, and the power variables (fi, e,)
of port 2 with dimension nz. Its constitutive relation is characterized by an (n, x IZJ
matrix M with coefficients in some field defining a vector space on the power
variables ; it is defined by Eq. (79) according to the representation in Fig. 15 :
,f2 = M,f,, el = Mke>. (79)
If the coefficients of the matrix M,Y are functions, the transformer is said to be
“modulated” and is denoted by MTF. The coefficients of the transformer are, for
instance, integers representing stoichiometric coefficients for chemical reactions’
models (14,15>, real numbers representing geometric parameters or real valued
functions of some energy and configuration variables for multibody systems (18).

+ (M)TF -+
fl f2

FIG. IS. A multiport transformer.

Journal of the Franklin lnst~tute

948 Pcrgamon Press Ltd
 Humiltoniun Formuhion qf Network Dynamics

FIG. 16. A multiport gyrator

The causal constraints at the ports of a multiport transformer depend on the rank
properties of M,. But for a single bond transformer, they are analogous to the
constraints on a single bond which may be considered as a unit transformer.
They provide constraints at the ports of a weighted junction structure which are
analogous to the constraints on the ports of a simple junction structure.
Finally, the complete network interconnecting capacitive elements, i.e. elemental
systems of different physical domains, may contain additionally some gyrative
element, denoted by GY and represented in Fig. 16. The complete network is
called : generalized ,junction structure. The constitutive relation of a gyrator is
defined by a skew symmetric matrix J, according to Eq. (SO) :

,f‘ = J,e, (80)

where the order of J., is equal to the dimension of the port of the multiport gyrator.
Again the causal constraints at its ports depend on the rank properties of J, but
they may be precised in the case of a two-port gyrator (necessarily of rank 2) : the
two causality assignments are depicted in Fig. 17. The causal constraints at the
port of a generalized junction structure reflect the invariant algebraic relations on
the power variables at the ports in the same way as for weighted junction structures.
Also, it has an additional interpretation if one connects the ports of a capacitor of
appropriate dimension to its ports, thus obtaining a complete (autonomous) bond
graph model. Then a causal restriction on the efforts at the ports implies the
restriction of the state space to a proper subset of R”, where n is the number of
energy variables. Indeed each causal constraint corresponds to some linear relation
on the efforts

and according to the definition of the effort variables at the ports of an energy
storage element this relation also induces a constraint on the energy variables :

-j(M)GY :&

//(M)GY : Jx
(b)
FIG. 17. Causal constraints on a two-port gyrator.

Vol 329. No. 5. pp. 923.966. 1992

Printed in Great Britain 949
B. M. Maschke et al.

(82)

where H(.Y) denotes the energy function of the n-port energy storage element.
In an analogous way, causal restrictions on the flows at the ports connecting the
generalized junction structure and the energy storage elements induce a restriction
on the time variation of the energy variables and hence on the dynamics of the
system which will be addressed in the following section.

E.xample 1.5. (Bond graph model of an electrical circuit.) Consider again the
electrical circuit of Example 2 and represented in Fig. 1. A generalized bond graph
model of this circuit is depicted in Fig. 18a. It contains three one-port energy
storage elements : two of them represent the inductances L, and Lz, and the third
one represents the capacitor C. The topology of the electrical circuit is realized by
a simple junction structure constituted by a l-junction with the bonds connected
according to Fig. I I. The symplectic gyrator indicates the coupling between the
magnetic and electrical domains (as in Example 12). The simple junction structure
together with the symplectic gyrator forms the generalized junction structure con-
necting the three energy storage elements in the complete generalized bond graph
model of the electrical circuit. The causally augmented bond graph depicted in Fig.
ISa shows that it is possible to assign integral causality to all energy storage
elements: the efforts at their ports are independent. As the energy variables are
related one-to-one to the effort variables through the constitutive relation of the
energy storage elements (here even linear relations), the dimension of the state-
space is three. Considering now the flows at the ports of the generalized junction
structure, the causally augmented bond graph of Fig. 18b shows the maximum
number ofindependent flows : at most two causal strokes may be directed outwards.
This corresponds to a constraint on the flows which is here quite simple: the
common flow expressed by the l-junction induces the following relation on the
generalized flows :

4, = p, = -I’? = _6;?. (83)

This relation indeed restricts the time variation of the energy variable to remain
within a plane defined by Eq. (84)

4,+42 = K, (84)
where K is some real number.

E.xamp/e 16. (Bond graph model of a one-dimensional mechanical system.)

Consider now the one-dimensional mechanical system of Example 6 and rep-
resented in Fig. 2. Its generalized bond graph model is analogous to Fig. 18 by
substituting the momentap, andpz for the magnetic flux linkages, the displacement
q, z of the spring for the electrical charge, the masses m, and tnn2for the inductances
and finally the compliance C for the capacitance. The simple junction structure
represents the force balance (and dually the kinematic constraints) in the system.
The symplectic gyrator indicates the coupling between the elastic and the kinetic

950
 Hamiltonian Formulation of Network Dynamics

___T____~________- Q;y:_ J”““__

(4 (b)

c+oK+1-$
- JVP

=o

++Odl
-Tg ww
Jg:c+io
1 \,s: -J

$$+o 31
-Iis
“1 = -v2

vv
-J

(4
FIG. 18. Bond graph models of the dynamics of an electrical circuit.

domains. The causal strokes in Fig. 18 indicate in the same way that the dimension
of the state space is three, but that the time variation of the state variables is
restricted to the force balance relation :

4, = F, = -F> = -&, (85)

where F, and F2 denote the forces exerted on the masses.

Example 17. (Bond graph model of a rigid body in free motion.) The model of

Vol. 329, No. 5, pp. 923-966, 1992

Printed in Great Britain 951
B. hf. Mmclzkc et al.

PY J : pz J ww

py : MTF MGY: pz
J +
~I-TM~W--TO
J JYV PX 3
:MGY td

(4

FIG. 19. Bond graph mod& of rigid body dynamics.

a rigid body in free motion is depicted in Fig. 19. The three one-port energy storage
elements represent the three components of the rotational kinetic energy of the
body along its principal axis with the projections p,, pv, p_ of the total angular
momentum on the principal axes as energy variables. The generalized junction
structure of Fig. 19 connected to the three capacitors provides the bond graph
realization of Euler’s equations (30). It is different from the “Eulerian junction
 Hamiltonian Formulation of Network Dynamics

structure” proposed by Karnopp (49) which is a nonlinear element in the velocities,

whereas here it is modulated by the momenta. The causality assignment of Fig.
19a shows that the three efforts at the ports of the generalized junction structure
are independent and hence the state space has dimension three. Figure 19b gives
at most two independent flows at the ports of the energy storage elements; thus
the variations of the energy variables are constrained on the so-called “invariant
plane” (2) defined by Eq. (86) :

3.2. The dynamics of‘ the energy tiariables as a Hamiltonian system

Here we are concerned with the analytical formulation of the dynamics associated
with the energy variables of a complete bond graph model : an n-port energy storage
element connected to a generalized junction structure. For the sake of simplicity
the following assumption is made on the causality at the ports of the generalized
junction structure: it admits flow causality at all its ports. According to the pre-
ceding section, this means that the efforts at the ports are independent. Hence the
generalized junction structure is equivalent at its ports to a modulated gyrator
(with constitutive matrix equal to minus the junction structure matrix J, due to
the power bond orientation according to Fig. 20a) with constitutive equation (80).
Furthermore the assumption on the independency of the efforts at the ports of the
generalized junction structure means that the energy variables of the connected
energy storage element are independent, which implies that the state space is simply
BY’.
If the causal assumption is not satisfied, then the space of the energy variables
(i.e. the state space) is restricted to a proper subset of R”, but the results presented
in this section remain valid provided that one restrict the maps defined hereafter
to this proper subset of R” induced by the causal constraints on the efforts at the
ports of the generalized junction structure (27).
Let us first interpret the constitutive relations of an energy storage element in
differential geometric terms. The energy function H(x) of the energy storage element
defines a smooth real-valued function on R”, the space of the energy variables X.
Identifying the tangent bundles 71w” of R” with I&!”x [w”and the cotangent bundle
T*R” with 58”x (0X”)*,the power variables at the ports of the energy storage element
may be interpreted as follows. The flow variable f, describing the time variation _t
of the energy variable is an element of the tangent space T,R” at the state x. The
effort variable e is the differential dH(x) of the energy function H(x) at X, thus is
a cotangent vector belonging to T\*R”. Then the constitutive relation of an energy
storage element defined in Eq. (65) may be defined as the local definition of the
application given in Diag. 2.
In the same way the relations on the power variables at the ports of the gener-
alized junction structure may be interpreted in differential geometric terms. Taking
into account the assumption on the causality restriction at its ports, the generalized
junction structure induces algebraic relations on the power variables at its ports
equivalent to the constitutive relations of a gyrator which may be modulated (see
Eq. 80). One may again identify the flow variables at its ports with an element of

Vol. 329. No. 5, pp. 923-966, 1992

Printed in Great Bntam 953
B. M. Muschkr et al.

P = dH.
H(x):am Qh+ [May: -J(x)
f =;, -1

(4

(c)

FIG. 20. Bond graph interpretation of the Hamiltonian reduction procedure.

the tangent space T,[W” at x and the effort variables as an element of the cotangent
space 7’prWnat x. Then the causal relations at the ports of the generalized junction
structure correspond to Diag. 3, where J, denotes the junction structure matrix.
This diagram would correspond to the mapping from one-forms to Hamiltonian
vector fields defined by a Poisson bracket if the junction structure matrix J, satisfies
the constitutive assumptions of structure matrices of Poisson structures. These

dH
e: IS”- T*(R”)

Id
Id: T(R”) - T(R”)

fx I- fx = ix
DIAG. 2.

Journal of thr Frenkhn Inamuw

954 Pergamon Press Ltd
 Hamiltoninn Formulation qfNetwork Dynumics

DIAG. 3.

constitutive assumptions are given by Eqs (11) and (12) and will now be discussed
with respect to the constitutive assumptions on junction structure matrices (or
gyrators’ constitutive relations). Firstly in order to ensure the power continuity of
the generalized junction structure, the matrix J, has to be skew-symmetric (i.e.
antisymmetric) (13); this is exactly the condition (11). On the other hand, the
Jacobi conditions (12) on a structure matrix (13) are usually not required for a
(generalized) junction structure matrix or the constitutive matrix of a gyrator. For
Hamiltonian systems, they play an essential role (see Section II) since they are
necessary and sufficient conditions in order to ensure that an arbitrary choice of
a Hamiltonian function induces a vector field which leaves the Poisson structure, i.e.
the geometric structure, of the system invariant. Hence in an analogous way for
networks they are necessary and sufficient conditions to ensure that one may choose
arbitrary energy functions for the energy storage elements independently of the
generalized junction structure. It is interesting to note moreover that as a conse-
quence of the Jacobi conditions, the invariants on the flow variables at the ports
of the generalized junction structure may be related one-to-one to invariants on
the energy variables whatever the actual energy function is. Physically this may be
explained by the very general derivation of power continuity from the conservation
of physical quantities like the electrical charge or the kinetic momentum (11, chaps
4 and 5). In terms of Hamiltonian systems, the Jacobi conditions allow to express
the degeneracy of the Poisson structure (corresponding to invariants on the flow
variables) as symmetries on the Hamiltonian function (defined on the state vari-
ables) of some embedding standard Hamiltonian systems as recalled in Section II.
Now we may consider again the complete bond graph model with causality assigned
according to the assumptions on the generalized junction structure and represented
on Fig. 20a. Assembling the diagrams representing the maps defined above, one
obtains the map represented in Diag. 4. This map is the differential geometric
interpretation of the usual construction of the dynamics of bond graph models
using causal paths in a generalized junction structure. Moreover, if the junction
structure matrix verifies the skew-symmetry and the Jacobi conditions (see Eq.
(12)) then the dynamics corresponds to a Hamiltonian system with Hamiltonian
function equal to the energy function H(x) on the Poisson manifold admitting the
junction structure matrix as structure matrix of its Poisson structure.

dH Jx
R” __, T*(R”) __* 7-(&l”)

x - dH, - &,

DIAG.4.

Vol. 329, No. 5, pp. 923-966, 1992

Pnnted in Great Britain 955
B. M. Maschke et al.

Exumple 18. (Dynamics of an electrical circuit.) Consider the circuit of Example

2 and its bond graph representation in Fig. 18a. The array of three one-port energy
storage elements may be interpreted as one three-port energy storage element with
energy function :

(87)

with Q the charge of the capacitor and $,, 4z the magnetic flux linkages of the
inductors. The energy function is the sum of the energy associated with each one-
port energy storage element. The constitutive relation of the three-port capacitor
is then linear and diagonal :

The causal relations at the ports of the junction structure in Fig. 1Xa induces the
following relation :

The matrix defining the linear inputtoutput relation at the ports of the junction
structure is the junction structure matrix. The relation (89) corresponds to the
relations induced by the fundamental loop and cutset matrices associated with the
tree in Fig. 14. Furthermore the junction structure matrix is skew-symmetric and
linear (and thus it satisfies necessarily the Jacobi identities). Thus it defines a
Poisson structure on R’. The combination of Eqs (87) (88) and (89) according to
Diag. 4, defines the dynamics of the electrical circuit as Hamiltonian dynamics on
R3 endowed with the Poisson structure defined by the junction structure matrix,
with the sum of the electrical and magnetic energy functions as Hamiltonian
function :

(90)

One recovers the dynamics presented in Example 2 ; however it was now deduced
using the causal procedure on the causally augmented bond graph model (12,42).

Exumple 19. (Dynamics of a mechanical system.) The one-dimensional mech-

 Humiltoniun Formulation ofNetwork Dynamics

anical system of Example 6 has a bond graph model analogous to Fig. 18a (see
Example 16). Hence its dynamics may also be expressed as a Hamiltonian system.
Its Hamiltonian function is the sum of the elastic potential energy of the spring
and the kinetic energies of the two masses :

This energy function implies the following constitutive relation at the ports of the
three-port energy storage element :

aH
kx,z
ax,>
aH PI
Fe, ml
e= 2,
I
=
ap, (92)
ii 2’2 aH P2
m2 ,
JPPZ

The generalized junction structure induces the following causal relations at the
ports of the junction structure :

.f=(jy) = (_r -i je. (93)

Combining Eqs (92) and (93) one recovers the Hamiltonian dynamics :

(3H
aLJI (94)
aH
c app2

Example 20. (Rigid body dynamics.) Consider again the rigid body in free motion
(Example 3) and its bond graph model in Fig. 19a. The array of the three one-port
energy storage elements may be grouped into one three-port energy storage element.
Its energy variable is simply the angular momentum p whose projections pv, p!, pz
on the three principal axes of the rigid body are the energy variables of the one-
port energy storage elements. The energy function is the rotational kinetic energy
of the rigid body and is equal to the sum of its components on the three principal
axes :

H(p, , p,, ) p_) = fi +

2J,y 2J,.
4; + p’
2J,.
(95)

The causal relations at the ports of the generalized junction structure are :

Vol. 329. No. 5, pp 923-966, 1992

Printed in Great Br~lam 957
B. M. Muschke et al.

(96)

The junction structure matrix of the bond graph model in Fig. 19a corresponds
exactly to the structure matrix of the Poisson structure of the Euler equations (see
Example 3).
The constitutive relations of the three-port energy storage element with energy
function H(I-‘) give :
r- -
PY
J,
Ei,
e=dH,=
P,
j- I ’ .f= p,, . (97)
P.- ii dz
-Jr,
Combining Eqs (96) and (97) according to Diag. 4, i.e. following the causal strokes
in the bond graph model, one recovers the Hamiltonian dynamics defined by the
Euler equations :

P\
7,

P?
J, (98)
P-
Jz

3.3. Relation M’ith stand& Humiltoniun systems und their bond graph realization
Now we shall discuss the relation of the dynamics of the energy variables of a
bond graph model with associated reduced and embedding standard Hamiltonian
systems. The discussion focuses on their bond graph realization, and special atten-
tion will be paid to the choice of coordinates of each representation.
First, the dynamics of the energy variables is discussed with respect to the reduced
standard Hamiltonian system obtained through the reduction by the dynamical
invariants. Its bond graph realization is based upon the analytical decomposition
of the generalized junction structure (13) : it makes appear explicitly its structure
and identifies the canonical variables in the decomposed bond graph. Secondly,
the bond graph realization of an embedding standard Hamiltonian system will be
obtained by the addition of some energy storage element with an energy function
which is identically equal to zero. Again, the associated symplectic structure and
canonical coordinates will be identified with some elements and flow variables of
the bond graph realization.

Journal dthc Frankhn Imt~tute

958 Pergamon Prw Ltd
 Hamiltonian Formulation of Network Dynamics

3.3.1. The reducedstandard Hamiltonian system. Let us consider the Hamiltonian

system describing the dynamics of the general bond graph model of Fig. 20a. In
order to reduce this system through its invariants, let us first consider its expression
in canonical coordinates following Section II. According to Theorem II, one knows
that there exists (locally) a change of coordinates denoted by 4 which provides
canonical coordinates :

444 = (4>/A~), (99)

where q and p are in R” and r is in [w’if 2n is the rank of the Poisson structure and
I is the dimension of its kernel.
In these coordinates, the structure matrix J, of the Poisson structure takes the
form :

0 4 0

Consequently
J(q,p, r) =
i
-I,
0
0
0 1
0 .
0

the structure matrix is split into a symplectic matrix of rank 2n and

(100)

a zero matrix corresponding to its kernel.

Moreover the change of coordinates 4 locally induces the two following adjoint
maps : the tangent map (& ‘)*_Yand the cotangent map (& ‘)z on the tangent and
cotangent bundles of R2”+‘. In coordinates, (& ‘)*.Ymay be represented by a matrix
r, (the subscript indicates that this matrix in general is a function of x) ; while the
adjoint map (de’): is represented by the transpose T_: of the real-valued matrix
r,. The matrix TX defines the analytical decomposition of the generalized junction
structure (13). The corresponding bond graph model is shown in Fig. 20b: the
generalized junction structure is decomposed into a modulated multiport trans-
former with constitutive matrix i”, relating the energy storage elements to a 2n-
port (symplectic) gyrator with constitutive matrix zymp :

(101)

and a zero-flow source of dimension 1.

This decomposition of the bond graph model was obtained in an analytical way
and may or may not coincide with the junction structure obtained from physical
considerations. It makes explicit the causal constraints on the flows at the ports of
the generalized junction structure in the form of the zero-flow source of dimension
1. According to the definition of the transformation matrix T,, the generalized
flows at the ports of the zero-flow source are the time-variation of the redundant
variables, which obviously obey the dynamic equation :

J = 0. (102)
In the same way the generalized flow variables at the ports of the 2n-port symplectic
gyrator correspond to the time-variations 4 and b of the canonical coordinates of
the reduced system. Thus they obey the constitutive relation :

Vol. 329, No. 5,pp. 923-966, 1992

Pnnted in Great Brian 959
B. M. Muschke et al.

-
aq
aH

(I ii
4 =
p
JSPV
(103)
n dH’
dP
Finally, one may observe that considering the energy flows, only the power bonds
at the ports of the symplectic gyrator contribute to the variation of the energy in
the energy storage elements. Indeed the power bonds with the flow variables being
the variation r: of the redundant variables convey a power equal to zero.

Example 21. Reconsider the electrical circuit of Fig. I and its bond graph model
in Fig. ISa. The analytically decomposed generalized junction structure, defined
by the change of variables (51) is shown in Fig. 18~. Compared to the physical
junction structure in Fig. 18a, a zero-flow source was added and a junction structure
relating it to the energy storage elements (they are denoted by the bold bonds in
Fig. 18~). The flow variable at the port of the zero flow source is
0 = r: = ;(c, fC2) = i(& +&). (104)
The zero-flow source shows explicitly the conservation of the total magnetic flux.
The flow variables at the ports of the symplectic gyrator are

4 = i, = 0, p = $(z\~-z~,) = {(d;2+&,). (103

The symplectic gyrator induces the dynamics of the reduced Hamiltonian system :
here it is a.linear harmonic oscillator, This reduced Hamiltonian system at the
ports of the symplectic gyrator is exactly the system obtained by Bernstein and
Lieberman (7). The change of coordinates performed in this example is the
alternative change of coordinates presented in (7) and leaving the capacitors’
charges invariant; they call the redundant variables r “latent” variables and the
canonical coordinates of the reduced system “active variables”. Summarizing. the
analytical decomposition of the generalized junction structure corresponds, for the
electrical circuit, to the addition of a simple junction structure and a zero-flow
source to the original physical junction structure. This additional junction structure
is necessary in order to define a regular change of coordinates, but is redundant in
the sense that its power bonds convey a power identically equal to zero. Moreover,
the generalized bond graph formalism uses the symplectic gyrator as a physical
coupling element of the network : at its ports one reads directly the reduced
Hamiltonian dynamics. This explains why one may define a topological reduction
procedure of the dynamics of electrical circuits as in (7).

Esumpk~ 22. Consider now the dynamics of the rigid body treated in Examples
3 and 17 and its bond graph representation in Fig. 19a. The change of variables
(60) would lead in the same way as for the electrical circuit, to the decomposition
of the junction structure into a two-port symplectic gyrator and a zero-flow source.
At the ports of the symplectic gyrator one recovers the reduced standard Ham-
iltonian dynamics and the zero-flow source represents explicitly the constraint on
the generalized flows :
 Hamiltonian Formulation of Network Dynamics

(106)
The junction structure matrix verifies the Jacobi conditions, hence the constraints
on the flow variables imply the existence of some invariant on the energy variables
which is here the conservation of the total momentum :

(107)
where k is some constant.
In contrast to the electrical circuit models, the analytical decomposition destroys
the original Eulerian junction structure realization. Furthermore, the analytical
decomposition of the generalized junction structure implies the use of varying
junction structures according to the singularity of the chosen coordinate trans-
formation. For instance, the coordinate transformation defined in (60) gives the
following singularity on the transformers moduli :

p,‘+pf = 0. (108)
Three different decomposed junction structures, obtained by circular permutation
of pl, .LQ,pZ are necessary to cover the entire state space. This disadvantage of
the analytical canonical decomposition of generalized junction structure may be
avoided using a graphical transformation. This transformation is based on the
“partial dualization” of the junction structure (50). It consists in making symplectic
gyrators appear at some ports of the junction structure by dualization, i.e. exchang-
ing the flows and effort variables at these ports, and minimizing the number of the
gyrators remaining in the junction structure. In the case of the Eulerian junction
structure (Fig. 19a), a partial dualization leads to the junction structure in Fig.
19~. It is obtained by dualization of the port attached to the energy storage element
representing the kinetic energy along the x-axis of the rigid body. Then the zero
junction connected to this port and the gyrators attached to this junction are
dualized into a l-junction and two transformers, leaving one gyrator in the junction
structure. In general the minimal sets of gyrators remaining in the junction structure
after partial dualization at its ports are called “essential” (SO). At the ports of the
symplectic gyrator one may again read the reduced standard Hamiltonian dynam-
ics. But the main advantage of the dualization procedure with respect to the
analytical decomposition is that the reduced Hamiltonian system is explicit graphi-
cally without generating any singularity on the remaining dualized junction
structure.

3.3.2. The embedding standard Hamiltonian system. Reconsider the general bond
graph model in Fig. 20a and its dynamics described by (15). In the previous
paragraph we have seen that the regular change of coordinates into canonical
coordinates leads to the analytical decomposition of the junction structure into a
symplectic gyrator (corresponding to the reduced dynamics) and makes appear a
zero-flow source representing the singularity of the junction structure. However,
according to Section II, one may further require the regularity of the geo-
metric structure of the dynamical system considering the embedding standard
Hamiltonian system (50).

Vol. 329, No. 5, pp. 923-966, 1992

Printed in Great Britain 961
B. M. Maschke et al.

Its bond graph realization arises from the decomposed junction structure in Fig.
20b ; the embedding system is defined in canonical coordinates (q,p, r) associated
with the Poisson structure. In order to regularize the geometric structure of the
system, the additional coordinate s is introduced, conjugated to variable Y (see Eq.
45). In bond graph terms, the generalized junction structure of the embedding
system is realized by introducing a bond carrying the flow variable S and a sym-
plectic gyrator of dimension 21 representing the conjugacy of S with i (see Fig.
20~). Hence the augmented generalized junction structure is now regular. On the
extended state space the Hamiltonian function of the embedding system (42) is :

fi(q, p, s, VI = H(q, p, y) + O(s), (10%

where O(s) represents the function of s equal to zero.

Equation (109) simply expresses that the Hamiltonian function His independent
of s, the vector of the conserved quantities of the embedding system. Thus the
energy associated with s is equal to zero : this is represented by an energy storage
element attached to the bond carrying S with an energy function equal to zero (see
Fig. 20~). Now the dynamics of the canonical coordinates (q,p,s,r) reads at the
ports of the two symplectic gyrators. At the ports of the symplectic gyrator of
dimension 2n one reads the dynamics of the reduced Hamiltonian system con-
sidering (109) :

In the same way, at the ports of the symplectic gyrator of dimension 21, one reads
the dynamics of the canonical variables s and Y.The time variation of the redundant
variables Y,becomes :

(111)

The invariance of the vector of redundant variable r is no longer expressed by a

zero-flow source (as in Fig. 20b) but by the degeneracy of an energy storage element
(Fig. 20~) which has an energy function equal to zero and thus constrains the
efforts at its ports to remain equal to zero. In network terms this means that this
storage element is equivalent at its port to a zero-effort source ; combined with the
symplectic gyrator this results again in an equivalent zero-flow source. The flow
variables at the ports of the energy storage elements are not causally constrained
by the extended generalized junction structure which is regular, but by the non-
invertibility of the constitutive relation (65) of the energy storage with energy
function equal to zero. The dynamics of the conserved quantities s is also read at
the ports of the symplectic gyrator of dimension 21:

(112)

962
 Humiltonian Formulation of Network Dynamics

Thereby the efforts aH(x)/&, conjugated to the zero-flow appearing with the ana-
lytical decomposition of a generalized junction structure, may be interpreted as
the time variation Sj of the conserved quantities of the embedding standard
Hamiltonian system.

Example 23. The bond graph representation of the embedding system associated
with the electrical circuit in Fig. 18a is shown in Fig. 18d. It was deduced from the
decomposed junction structure in Fig. 1%~by replacing the flow source by a gyrator
and an energy storage element with an energy function equal to zero. At the ports
of the symplectic gyrator one reads :

.$ = -(i, +i*), (21,+u,) = i = P,,= 0, (113)

where e,sdenotes the effort at the port of the degenerate energy storage element.
Thus the degeneracy of the additional energy storage element ensures the con-
straint on the voltages. The current type variable - (i, + iJ conjugated to the
redundant variable may be interpreted as the time variation of the conserved
quantity of the embedding standard Hamiltonian system. It may further be noticed
that the embedding system does not have a realization in electrical circuit terms as
the simple junction structure in Fig. 18c, d is not graphical (45).

Example 24. Take again the one-dimensional mechanical system of Example 6,

which has a bond graph representation analogous to the electrical circuit, and
consider the bond graph representation of its embedding standard Hamiltonian
system in Fig. 18b. Then at the ports of the additional symplectic gyrator, one
reads again the dynamics of the redundant variable r = (p, +p3), i.e. the total
momentum of the system and of its conjugated conserved quantity :

s = -(c, +c2), (0, +J$) = i = L’, = 0. (114)

The conserved quantity may be interpreted as the position (up to some multi-
plicative constant) of the masses.

IV. Conclusion

The structural relations between network representations of systems not con-

taining resistive (or irreversible) phenomena (such as LC electrical circuits or
mechanical systems without friction) and the Hamiltonian formalism of analytical
mechanics were investigated.
To this aim, generalized Hamiltonian systems were considered, i.e. systems
defined on degenerate Poisson manifolds which appear in the context of reduction
of standard Hamiltonian systems with symmetry. This generalization allows in
particular the handling of state spaces of odd dimension and the explication of the
invariants of the system in terms of the geometric structure, i.e. the Poisson
structure. One interesting consequence of the use of Poisson structures is that it
implies the use of dual sets of variables.
The bond graph notation was used for the network representation of systems
and the generalized bond graph formalism allowed to make the relation between

Vol 329. No. 5. pp. 923-966, 1992

Printed m Great Bntam 963
B. M. Muschke et al.

the network and Hamiltonian formalisms. Indeed its main feature is to symmetrize
all energetic network elements in a single energy storage element and accordingly to
explicate the coupling between different physical domains by a unit gyrator element,
thus endowing the network including the gyrators (and called generalized junction
structure) with a gyrator-type constitutive relation at its ports.
Then the relation between the network and the Hamiltonian formalism is made
by matching the energy function associated with the energy storage elements of the
network with the Hamiltonian function and the generalized junction structure
relations at its ports with the local definition of a Poisson structure. It was proved
that the dynamic equations associated with the energy variables (of the energy
storage element) is Hamiltonian and that the topological (or graphical) con-
struction of these dynamics matches exactly the analytical definition of Hamiltonian
dynamics.
The crucial point of the relation of the network and Hamiltonian formalisms is
the representation of the dynamic invariants. Indeed the (general) Poisson structure
explicates the invariants of the system as the elements of its kernel. These may be
related with the algebraic and topologic constraints induced by the generalized
network (the generalized junction structure) at its port variables.
Furthermore, a network interpretation of the relation with the embedding and
reduced standard Hamiltonian systems was given. The reduced Hamiltonian system
was shown to correspond to the generalized flows at the ports of the symplectic
gyrators, appearing as coupling elements between the different physical domains
or by the analytical decomposition of the generalized junction structure. The
embedding system was constructed using a degenerate energy storage element with
energy function equal to zero and the symmetry variables of the embedding system
as energy variable.
In summary, the consideration of generalized Hamiltonian systems defined with
respect to degenerate Poisson structure is essential for an intrinsic analytical rep-
resentation of the main features of network representations: the duality of port
variables and the invariants intrinsically given by the network topology.

References

(1) V. 1. Arnold, “Mathematical Methods of Classical Mechanics”, Springer, Berlin, 1978

(Translation from the 1974 Russian edition).
(2) R. Abraham and J. E. Marsden, “Foundations of Mechanics”, Second Edn, Benjamin
Cummings, Reading, MA, 1978.
(3) G. Marmo, E. Saletan, A. Simoni and B. Vitale, “Dynamical Systems : a Differential-
geometric Approach to Symmetry and Reduction”, Wiley, New York, 1985.
(4) G. Galavotti, “The Elements of Mechanics”, Springer, Berlin, 1983.
(5) P. Libermann and C. M. Marle, “Symplectic Geometry and Analytical Mechanics”,
Reidel, Dordrecht, 1987.
(6) P. S. Kirshnaprasad and S. E. Marsden, “Hamiltonian structures and stability for
rigid bodies with flexible attachments”, Archs rution. Mech. Anal., Vol. 98, pp. 71-
93. 1987.
(7) G. M. Bernstein and M. A. Lieberman, “A method for obtaining a canonical Ham-
 Humiltonian Formubtion of Network Dynamics

iltonian for nonlinear LC circuits”, Trans. IEEE Circuits Syst., Vol. CAS-36, pp.
41 I-420, 1989.
(8) R. W. Brockett, “Control theory and analytical mechanics”, in “Geometric Control
Theory” (Edited by C. Martin and R. Herman), Vol. 7 of “Lie Groups: History.
Frontiers and Applications”, pp. l-46, Mathematical Science Press, Brookline,
1977.
(9) A. J. van der Schaft, “System Theoretic Descriptions of Physical Systems”, CWI Tracts
3. CWI, Amsterdam, The Netherlands, 1984.
(10) A. J. van der Schaft, “System theory and mechanics” in “Three Decades of Math-
ematical Systems Theory”, Lecture Notes in Control and Inf. Sci., Vol. 135, pp.
4266452, Springer, Berlin, 1989.
(11) H. M. Paynter, “Analysis and Design of Engineering Systems”, MIT Press, Cambridge,
MA, 1961.
(12) D. C. Karnopp and R. C. Rosenberg, “Systems Dynamics : a Unified Approach,
Wiley, New York, 1975.
(13) P. C. Breedveld. “Physical systems theory in terms of bond graphs”, Ph.D. Thesis,
University of Twente, The Netherlands, 1984.
(14) G. F. Oster and A. S. Perelson, “Chemical reaction dynamics. Part I : geometrical
structure”, Archs ration. Mech. Anal., Vol. 55, pp. 236273, 1974.
(15) A. S. Perelson and G. F. Oster, “Chemical reaction dynamics. Part 2: reaction
networks”, Archs ration. Mech. Anal., Vol. 55, pp. 3 l-98, 1975.
(16) J. J. van Dixhoorn, “The use of network graphs and bond graphs in 3D-mechanical
models of motor cars and unbalance”, in “Physical Structure in Systems Theory”
(Edited by J. J. van Dixhoorn and F. J. Evans), pp. 83-l 10, Academic Press, London,
1974.
(17) R. R. Allen, Multiport representation of inertia properties of kinematic mechanisms”,
J. Franklin Inst., Vol. 308, pp. 2355253, 1979.
(18) A. M. Bos and M. J. L. Tiernego. “Formula manipulation in the bond graph modelling
and simulation of large systems”, J. Franklin Inst., Vol. 319, pp. 51-65. 1985.
(19) F. H. Branin Jr, “The network concept as unifying principle in engineering and the
physical sciences”, in “Problem Analysis in Science and Engineering” (Edited by F.
H. Branin and K. Huseyin), pp. 41-l 11, Academic Press, New York, 1977.
(20) H. M. Trent, “Isomorphisms between oriented linear graphs and lumped physical
systems”, J. Acoust. Sot. Am., Vol. 27, pp. 500-527, 1954.
(21) F. J. Evans and J. J. van Dixhoorn, “Towards more physical structure in systems
theory”, in “Physical Structure in System Theory” (Edited by J. J. van Dixhoorn
and F. J. Evans), pp. 83-l 10, Academic Press, London, 1974.
(22) P. C. Breedveld, “Multibond graph elements in physical systems theory”, J. Franklin
Inst., Vol. 319.. pp. l--36, 1985.
(23) D. L. Jones and F. J. Evans, “Variational analysis of electrical networks,” J. Franklin
Inst., Vol. 295, pp. 9-23, 1973.
(24) P. C. Breedveld, “Thermodynamic bond graphs: a new synthesis”, Int. J. Modell.
Simul., Vol. 1, pp. 57-61, 1981.
(25) P. C. Breedveld, “Thermodynamic bond graphs and the problem of thermal inertance”,
J. Franklin Inst., Vol. 314, pp. 15-40, 1982.
(26) S. Smale, “On the mathematical foundations of electrical circuit theory”, J. LX3
Geom., Vol. 7, pp. 193-210, 1972.
(27) B. M. Maschke, “Geometrical formulation of the bond graph dynamics with appli-
cation to mechanisms”, J. Franklin Inst., Vol. 328, pp. 723-740, 1991.
(28) S. Lie “Theorie der Transformationsgruppen”, Vol. 2, B. G. Teubner, Leipzig, 1890.

Vol. 329, No. 5, pp. 923-966. 1992

Prmted in Great Britain 965
B. M. Maschke et al.

(29) C. Carathkodory, “Calculus of Variations and Partial Differential Equations of the

First Order”, Vol. I, Holden Day, San Francisco, 1965.
(30) A. Weinstein, “The local structure of Poisson manifolds”, J. D@ Geom., Vol. 18, pp.
523-557, 1983.
(31) R. Hermann, “Gauge Fields and Cartan-Ehresmann Connections : Part A : Inter-
disciplinary Mathematics”, Vol. X, Mathematical Science Press, Brookline, 1975.
(32) A. Lichnerowicz, “Les vari&Cs de Poisson et leurs algkbres de Lie associkes”, J. Dif:
Geom., Vol. 12, pp. 253-300, 1977.
(33) D. C. Karnopp, “Lagrange’s equations for complex bond graph systems”, Truns.
ASME J. Dyn. Syst. Meus. Control, Vol. 110, pp. 306306, 1977.
(34) D. C. Karnopp, “Alternative causal patterns and equation formulation for dynamic
systems”, Trans. ASME J. Dyn. Syst. Meus. Control, Vol. 105, pp. 58-63, 1983.
(35) G. Recb, “VariCtCs symplectiques, varikti-s presque-complexes et systkmes dyna-
miques”, C. R. Acad. Sci. Paris, Vol. 235, pp. 776-718, 1952.
(36) E. Cartan, “LeCons sur les lnvariants IntCgraux,” Herman. Paris, 1922.
(37) P. J. Olver, “Applications of Lie Groups to Differential Equations”, Springer, New
York, 1986.
(38) V. I. Arnold, “The Hamiltonian nature of the Euler equations in the dynamics of a
rigid body and an ideal fluid”, Usp. Mut. Nuuk, Vol. 24, pp. 225-226, 1969 (in
Russian).
(39) E. T. Whittaker, “A Treatise on the Analytical Dynamics of Particals and Rigid
Bodies”, Cambridge University Press, Cambridge, 1937.
(40) J. E. Marsden and A. Weinstein, “Reduction of symplectic manifolds with symmetry,”
Rep. Math. Phys., Vol. 5, pp. 121-130, 1974.
(41) J. E. Marsden and A. Weinstein, “Coadjoint orbits, vertices and Clebsch variables for
incompressible fluids”, Physiccr 7D. pp. 305-332, 1983.
(42) N. Hogan and E. D. Fasse. “Conservation principles and bond graph junction struc-
ture”, Automated Modelling for Design, Trans. ASME DSC-Vol. 8, Winter Annual
Meeting of the ASME, Chicago, IL. 28 Nov.-2 Dec. 1988.
(43) C. Bidard, “Displaying Kirchhoff’s invariants in simple junction structure”, Proc. 5th
IMACS World Congress, Dublin, Ireland, 22-26 July 1991.
(44) P. C. Breedveld. “A systematic method to derive bond graph models”. Trans. 2nd
Eur. Simulation Congress, Antwerpen. Belgium, 1986.
(45) S. H. Birkett and P. H. Roe, “The mathematical foundation of bond graphs-l.
Algebraic theory; II. Duality; 111. Matroid theory: IV. Matrix representation and
causality”. J. Franklin Inst.. Vol. 326, pp. 329-350. 691-708, 1989; Vol. 327. pp.
87.-108, 109-128, 1990.
(46) A. S. Per&on and G. F. Oster, “Bond graphs and linear graphs”, J. Franklin Inst.,
Vol. 302, pp. 159-185. 1976.
(47) C. Bidard. “Bond graph procedure for the structural analysis of mechanisms-Part I:
kinematic junction structures and causality assignment”, Proc. 5th IFToMM Int.
Symp. Linkages and Computer Aided Design Methods, Bucharest, Rumania, 6-l I
July 1989.
(48) C. Bidard, “Kinematic structure of mechanisms : a bond graph approach”, J. Franklin
Inst., Vol. 328, pp. 9Oi-915, 1991.
(49) D. C. Karnopp, “Power conserving transformations : physical interpretations and
applications using bond graphs”, J. Franklin Inst.. Vol. 288, pp. 175-201, 1969.
(50) P. C. Breedveld, “Essential gyrators and equivalence rules for three-port junction
slructurcs”. J. Frunklin Inst., Vol. 3 18, pp. 253-273, 1984.

966